identity:
  role: Enterprise developer by day. AI runtime builder after hours.
  thesis: Five years shipping at scale taught me where things break.
          I build the layer that prevents it.
  approach: Build the thing that solves the problem. Not the thing that demos well.

A local-first thinking runtime for autonomous agents. Not a framework - a runtime.

Every LLM call gets a compiled phase packet: typed context filtered by phase, not dumped as a transcript. Planning sees durable anchors. Acting sees tool evidence. Response generation never sees skill instructions. Three independent safety layers catch what the model misses.

If a tool write fails, the response cannot claim it succeeded. If a fact changes, the old one is voided - not appended. Runs on a 3B model, fully local.

Providers → Ollama · LM Studio · llama.cpp · OpenAI · Anthropic · Groq · Gemini

Day job - Lead developer on enterprise business applications. End-to-end ownership: JavaScript, Backbone.js, Java Spring, Oracle - through to production deployment. Client refinement, feature intake, QA triage, release management.

Before that - Five years building on a BI platform used by over a million people. Reduced data retrieval times by 70%. Resolved 100+ security vulnerabilities. Shipped quarterly releases with zero deployment failures. Scale teaches you things a tutorial never will.

Won an internal AI Innovation Challenge building a RAG assistant over a large open dataset. That project started the thinking behind shovsOS.

Independent SaaS:

One of these hit 90% profit margin within 60 days. Sole developer. No marketing budget. No co-founder.

 

I write about the gaps between how AI systems are described and how they actually behave in production.

→ Medium